Subscribe now

Americans broadly favor responsible AI policies

With the rise of Generative AI, the desire for more transparent and responsible use of AI has also increased. Many in the AI community have called for greater accountability, including the labeling of AI-generated content, enhanced checks for bias and fairness, strengthened data security and privacy, and more explainable AI systems.1

Data from the 2023 Generative AI & American Society Survey show that Americans broadly support Responsible AI policies, with agreement ranging from 41% to 66%, depending on the specific policy. The proportion of respondents who are neutral — neither agreeing nor disagreeing — varies between 20% and 31%, while explicit disagreement remains low across all policies.

Measures to encourage Responsible AI are not without tradeoffs. These policies could increase the cost to use AI tools or slow AI innovation, blunting potential benefits. Moreover, integrating safeguards into AI tools to prevent the generation of harmful or offensive content implies that AI companies would have the authority to define what is considered “harmful” or “offensive,” potentially leading to censorship concerns.

These trade-offs were considered when gauging the public’s level of agreement with each facet of Responsible AI. For instance, the wording of the question pertaining to content safeguards is presented below. The complete phrasing for each Responsible AI question can be found at the end of this article or on this page.

Idea: AI tools generating content should have built-in safeguards to avoid creating harmful or offensive content.

Benefit: This could result in content that is more respectful and less likely to cause emotional harm.

Tradeoff: This could mean that companies define what is considered “harmful” or “offensive”. This might lead to censorship issues.

Responsible AI is already making its way into various U.S. policy proposals. Senator Chuck Schumer recently proposed his SAFE Innovation Framework for Generative AI at a speech at the Center for Strategic and International Studies. Schumer hopes to sponsor and pass his framework as Congressional legislation.

One proposal from the SAFE Innovation Framework involves measures to address bias, a topic that has been widely reported on. Schumer’s home state of New York has already passed legislation requiring AI systems used in hiring decisions to undergo annual audits for bias. Even when presented with tradeoffs like the delayed release of AI systems or less creative AI, 58% of Americans agreed that AI systems should undergo thorough checks for bias and fairness before launch, with just 8% in disagreement.

However, New York’s law does not mandate so-called “AI explainability,” the concept of decoding the specific inner workings of AI systems to understand how they make decisions. The idea of AI systems that explain their decision making received the support of 55% of Americans, even if it makes AI systems harder to build and increases costs to use them. Only 6% disagreed with the policy idea.

In June of 2023, U.S. Senator Michael Bennet, a Democrat who is active in artificial intelligence issues, urged major tech firms to impose labels on AI-generated content to help curb the dissemination of material designed to mislead users. The 2023 Generative AI & American Society Survey reveals that 63% of Americans agree with labeling AI-generated content, even considering potential tradeoffs like the unnecessarily stigmatizing educational or useful AI content. Public disagreement with labeling is at a mere 4%.

Others in the AI community have called out security vulnerabilities of new AI language systems, such as prompt injections. Nearly two-thirds of Americans (63%) support AI systems prioritizing user security and data privacy, even if it makes the AI system harder to use. Only 4% disagree.

Tech companies are promoting their Responsible AI practices

Major technology companies have already begun to tout their increased focus on Responsible AI. Microsoft CEO Satya Nadella and a group of Microsoft Executives underscored the company’s Responsible AI practices in their M365 Copilot announcement in May of this year. Just a week later Google CEO Sundar Pichai authored an op-ed in the Financial Times titled, “Building AI responsibly is the only race that really matters.” That same month OpenAI co-founder and CEO Sam Altman urged U.S. Congress to increase AI regulation in his Senate testimony.

In June, Microsoft, Google, OpenAI and four other major AI companies agreed to a series of Responsible AI practices proposed by the Biden administration. Earlier this month the White House announced that eight more companies had signed on.

Some remain skeptical that companies’ voluntary commitments will provide the necessary measures to reduce AI harms. Emily Bender, an AI researcher and outspoken critic of Big Tech, argued in July that effective legislation must be imposed from outside the tech industry. Another critic, AI expert and tech founder Gary Marcus, penned an op-ed in Time titled, “In the Rush to AI, We Can’t Afford to Trust Big Tech.”

Meanwhile, Senator Schumer’s SAFE Innovation Framework is moving forward. Elsewhere in Congress, a bipartisan group, led by Ted Lieu of California, has proposed additional regulation through a National Commission on Artificial Intelligence. Whether critics support the approach or not, Big Tech companies will likely be a part of the regulation decision-making process.

Share

Higher levels of education correlate with agreement on Responsible AI; Some remain neutral

There is moderate demographic variation in support for Responsible AI policies when agreement is averaged across the eight distinct areas of Responsible AI. Individuals with higher educational attainment tend to exhibit more support. Specifically, 68% of those holding a bachelor’s degree or higher express support for Responsible AI policies, contrasting with the 40% support among those without a high school diploma. However, this disparity largely stems from individuals with lower education levels expressing more neutrality toward the policies rather than outright disagreement. Those with less educational attainment are also less familiar with Generative AI technologies and so may not fully appreciate the implications of Responsible AI policies.

Similar trends are observed in other demographic categories. Older Americans and those with higher incomes are more inclined to support Responsible AI policies. Regarding ethnic differences, Black individuals are less likely to agree with the policies; but again this is mostly due to their high rates of neutrality rather than substantially higher rates of disagreement.

In fact, disagreement with Responsible AI is consistently low across all demographic groups, with disagreement ranging between 6% and 11% depending on the specific demographic considered.

A majority of Americans support AI antitrust regulation

The majority of Americans also support regulation to prevent AI market concentration, even if it slows the pace of AI innovation.

Market concentration has been a concern for some in the AI community. For instance, the AI NOW Institute submitted a public comment to the U.S. Federal Trade Commission after they invited input on regulating cloud computing, which serves as the infrastructure for most AI technology. The Institute wrote:

[The large datasets needed to create AI] deeply entrench the infrastructural and economic power of the few firms that retain control over the key components to building AI, with detrimental effects on competition in the AI industry. This also contributes to consumer injury in many forms, including harms to privacy and security, encouraging the spread of false and misleading information, perpetuating patterns of inequality and discrimination, harmful effects on workers, and environmental harms.

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about Responsible AI are shown below. For more information, please refer to the FAQ page.

Responsible AI questions

A total of 13 policy questions were presented in a grid format, grouped into three thematic categories. Each question allowed respondents to choose their level of agreement. The order of the 13 questions presented to respondents was randomized. Each policy question included an idea along with potential benefits and tradeoffs. The 8 Responsible AI questions were among this broader set of 13 policy questions.

Theme 1 - Openness and responsibility

The following are some ideas to make AI systems more open and responsible. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Question 1

Idea: AI systems should clearly explain how they make decisions.

Benefit: Clear AI explanations could help users make smarter choices.

Tradeoff: This could make AI systems harder to build so they might cost more to use.

Question 2

Idea: Companies should tell us how their AI models work and what data they use.

Benefit: This approach could foster trust and allow users to pick the AI systems they feel comfortable using.

Tradeoff: This could give away AI company secrets and let AI creators in other countries get ahead. In addition, knowing they have to disclose details might make companies less eager to innovate.

Question 3

Idea: AI companies should tell users when they're talking to an AI, not a person.

Benefit: Knowing they're talking to an AI could help users set clear expectations about the interaction.

Tradeoff: This could break the feel of a real chat and make AI systems less enjoyable or easy to use.

Question 4

Idea: Content created by AI should have a clear label saying it's AI-generated. AI-generated content includes images, videos, stories, and music.

Benefit: This could enable users to understand content origins and make informed decisions.

Tradeoff: People might engage with content less if they know it's made by AI even if the content is useful or entertaining. That's because it could seem less genuine or appealing.

Theme 2 - Changes to operations

The following are some ideas for potential changes to how Generative AI systems and companies operate. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Question 5

Idea: AI systems should undergo thorough checks for bias and fairness before their launch.

Benefit: If AI is fair, it could mean everyone gets an equal shot, especially when big decisions are made.

Tradeoff: This could delay the release of the AI system or make the system less creative.

Question 6

Idea: AI systems should be designed to prioritize user privacy and data security.

Benefit: This could help keep personal data out of the wrong hands.

Tradeoff: This might make the AI less useful or harder to use.

Question 7

Idea: AI tools generating content should have built-in safeguards to avoid creating harmful or offensive content.

Benefit: This could result in content that is more respectful and less likely to cause emotional harm.

Tradeoff: This could mean that companies define what is considered “harmful” or “offensive”. This might lead to censorship issues.

Theme 3 - Government regulation

The following are some ideas for how governments might handle Generative AI. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Question 8

Idea: The government shouldn't let AI assist in running crucial systems. These include power grids, transit, telephone lines, or the internet.

Benefit: By excluding AI, we might reduce risks linked to AI errors or cyber attacks in these systems.

Tradeoff: Lack of AI in these systems might reduce automation and increase user costs. For example, your power bill could be higher than if AI were used.

Market concentration question

Idea: Regulations should stop a few technology companies from controlling the entire AI market.

Benefit: This could boost competition. It could allow for more diverse AI solutions. It could also prevent one company from getting too powerful.

Tradeoff: This could slow down the pace of AI development and innovation.

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

Subscribe now

Americans support paying artists if their work is used in AI model training

The rise of Generative AI has placed creators — including photographers, authors, musicians, and other artists — squarely in the spotlight. New Generative AI systems are trained on large datasets, some of which contain copyrighted material. While many artists argue that such usage infringes on their copyrights, AI companies assert that their actions fall under the umbrella of U.S. fair use laws.1

Recent legal challenges,2 such as comedian Sarah Silverman and a group of authors suing OpenAI and Meta for copyright infringement, highlight this contentious issue. The probable outcome of these lawsuits remain uncertain. In July, a court asked artists in one copyright lawsuit against Midjourney and two other text-to-image platforms to provide more evidence for infringement before the case could move forward.

In exchange for Generative AI companies using copyrighted material during model training, some advocates in the AI community have suggested paying creators for the use of their work.

Setting aside the legal questions about using artists' creations for AI training, the 2023 Generative AI & American Society Survey shows that most Americans are in favor of compensating artists when their work is used in this way (or allowing them to opt-out).

A majority of Americans, 57%, agree with compensation for artists while only 8% of Americans disagree. Just under a quarter of Americans (23%) neither agree nor disagree.

In gauging public sentiment about paying artists or allowing them to opt out, the survey question included both the potential benefits and tradeoffs of such a policy. Americans’ agreement was solicited using the following question wording:

Idea: Artists, authors, and musicians should be paid if AI systems use their works to learn. Or they should have the choice to opt-out.

Benefit: Creators could get recognized and financially rewarded for their work.

Tradeoff: This could limit the diversity of AI training data, affecting the quality of AI content. It might also increase the cost of using AI systems.

While all demographic groups show at least moderate support, there was variation. Younger Americans and those under 30 were the least likely to support such a proposal as were low-income Americans (those making less than $30,000 per year), and Americans with low educational attainment.

Conversely, support was highest among those 30 and over and for those with incomes above $60,000. Among all demographic groups analyzed, those with a bachelor's degree or higher showed the highest level of support for the policy, at 71%.

Share

Americans also favor labeling AI content and more transparency about what data is used for AI model training

Creators encounter challenges in determining whether and how their works are used for AI training, partly because some companies, such as OpenAI, do not disclose their training data.3 The 2023 Generative AI & American Society Survey indicates that half of Americans think AI companies should be more transparent about how Generative AI models work and what data are used to train them. Another 30% neither agree nor disagree, while only 8% explicitly disagree.

Some Generative AI companies, such as OpenAI, cite strategic and competitive reasons as contributing factors for not disclosing specific details about the training and implementation of resource-intensive projects like GPT-4. The potential anti-competitive nature of revealing training data and model details was accounted for in the survey question wording:

Idea: Companies should tell us how their AI models work and what data they use.

Benefit: This approach could foster trust and allow users to pick the AI systems they feel comfortable using.

Tradeoff: This could give away AI company secrets and let AI creators in other countries get ahead. In addition, knowing they have to disclose details might make companies less eager to innovate.

Labeling AI-generated content has been proposed as a method to counter misinformation, however, it could also give consumers the option to abstain from engaging with content that raises ethical questions for them about the unauthorized use of creators' works. Some major content platforms, such as TikTok, have already begun enforcing such labeling policies. According to TikTok's updated guidelines, any content using AI to create a realistic depiction of a scene must be clearly labeled as AI-generated.

The 2023 Generative AI & American Society Survey found that a significant majority of Americans, 63%, support labeling AI-generated content, while a mere 4% oppose it. Another 23% neither agree nor disagree on the policy.

As I’ve discussed elsewhere, labeling AI content could risk unnecessarily stigmatizing content that is engaging, educational, or otherwise useful. However, this potential tradeoff was explicitly taken into account in the wording of the survey question used to gauge public sentiment.

Idea: Content created by AI should have a clear label saying it's AI-generated. AI-generated content includes images, videos, stories, and music.

Benefit: This could enable users to understand content origins and make informed decisions.

Tradeoff: People might engage with content less if they know it's made by AI even if the content is useful or entertaining. That's because it could seem less genuine or appealing.

Share

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about creator payment, AI transparency, and AI labeling are shown below. For more information, please refer to the FAQ page.

Paying creators question

The following are some ideas for potential changes to how Generative AI systems and companies operate. Each idea has benefits, but there could also be downsides, or "tradeoffs." We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Idea: Artists, authors, and musicians should be paid if AI systems use their works to learn. Or they should have the choice to opt-out.

Benefit: Creators could get recognized and financially rewarded for their work.

Tradeoff: This could limit the diversity of AI training data, affecting the quality of AI content. It might also increase the cost of using AI systems.

• Strongly agree

• Agree

• Neither agree nor disagree

• Disagree

• Strongly disagree

• Don’t know

AI-generated content labeling question

The following are some ideas to make AI systems more open and responsible. Each idea has benefits, but there could also be downsides, or "tradeoffs." We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Idea: Content created by AI should have a clear label saying it's AI-generated. AI-generated content includes images, videos, stories, and music.

Benefit: This could enable users to understand content origins and make informed decisions.

Tradeoff: People might engage with content less if they know it's made by AI even if the content is useful or entertaining. That's because it could seem less genuine or appealing.

• Strongly agree

• Agree

• Neither agree nor disagree

• Disagree

• Strongly disagree

• Don’t know

AI company transparency question

The following are some ideas to make AI systems more open and responsible. Each idea has benefits, but there could also be downsides, or "tradeoffs." We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Idea: Companies should tell us how their AI models work and what data they use.

Benefit: This approach could foster trust and allow users to pick the AI systems they feel comfortable using.

Tradeoff: This could give away AI company secrets and let AI creators in other countries get ahead. In addition, knowing they have to disclose details might make companies less eager to innovate.

• Strongly agree

• Agree

• Neither agree nor disagree

• Disagree

• Strongly disagree

• Don’t know

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

Subscribe now

Americans believe creative and tech jobs will be most harmed by Generative AI

The future of work has garnered significant attention as new Generative AI technologies demonstrate their capabilities in tasks previously considered uniquely human. Media outlets including The Washington Post, BBC News, Fast Company, Fox News, and Time Magazine have all explored the workforce implications of emerging Generative AI. A research paper from OpenAI estimated that 50% of jobs will have at least 50% of their tasks automated by AI language tools in the coming years.

However, there has been limited research on the views of everyday Americans regarding the impact of AI across job sectors. The 2023 Generative AI & American Society Survey shows that American perceptions on the impact of Generative AI across various job areas are diverse. No single job area was chosen by more than about 20% of respondents as the main area to either benefit from or be harmed by Generative AI. (Note that the timeframe provided in the survey questions specifically referred to “the next three years.”).

When it comes to which areas Americans think will be most negatively affected, 21% of respondents believe that creative professions like artists and writers will bear the brunt of Generative AI's impact. However, only 7% of those surveyed think that this field stands to benefit the most from the technology.

This is perhaps no surprise. Creative jobs have been at the center of early lawsuits against Generative AI companies, mainly focused around copyright infringement,1 with fears that artists’ work will be used against them.

Tech-related jobs, including programmers and data analysts, stood out as the most polarizing job area, receiving the highest percentage (21%) for “most likely to benefit” from Generative AI, while also ranking second highest (13%) for “most likely to be harmed.”

A larger percentage of Americans (9%) think that media-related jobs, such as game designers and 3D artists, will benefit more from Generative AI, compared to the 6% who believe these roles will be most harmed. Game design is an area that has shown recent progress integrating Generative AI technology (see video below).

Against the backdrop of ongoing strikes by writers and actors unions, with AI as a major point of contention, 5% of Americans believe that performing arts jobs will be the area most harmed by Generative AI.

In finance, teaching, and sales and marketing job areas, Americans held fairly balanced views. Between 5% to 7% of respondents believed that these job areas will both benefit and be harmed the most by Generative AI

Perceptions of Americans do not always align with research on job exposure to Generative AI. For instance, while Americans estimated limited impact of Generative AI to jobs in the finance field, the OpenAI paper cited earlier found that many finance jobs are among the most exposed to AI language tools like ChatGPT, (though note that exposure could mean either human augmentation, a benefit, or human replacement, a harm).

Notably, there's a large degree of uncertainty among Americans about Generative AI's potential effects on the job market. Nearly one-third of respondents expressed uncertainty about which job sector will gain the most from Generative AI, while around one-quarter were uncertain about which will be the most adversely affected.

Share

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about job areas are detailed below. For more information, please refer to the FAQ page.

The order of these two questions was randomized so that half of respondents saw the job benefit question first and the other half saw the job harm question first. In addition, each respondent saw the job areas presented in a random order.

Generative AI job benefit question

Think about the next three years. Which job area do you believe will benefit the most from Generative AI? For each job area we've given some example jobs, but there might be other jobs in the job area.

• Creative jobs (Artists, Writers, Designers)

• Media jobs (Game designers, Game programmers, 3D artists)

• Performing arts jobs (Movie and TV actors)

• Finance jobs (Accountants, Stock analysts, Traders)

• Tech jobs (Programmers, Data analysts, Engineers, IT)

• Teaching jobs (Teachers, Course creators)

• Legal jobs (Lawyers, Legal assistants)

• Health jobs (Doctors, Nurses)

• Sales and Marketing jobs (Sales representatives, Brand managers, Digital marketers)

• Other (please specify)

• I don’t know

Generative AI job harm question

Think about the next three years. Which job area do you believe will be most hurt by Generative AI? For each job area we've given some example jobs, but there might be other jobs in the job area.

• Creative jobs (Artists, Writers, Designers)

• Media jobs (Game designers, Game programmers, 3D artists)

• Performing arts jobs (Movie and TV actors)

• Finance jobs (Accountants, Stock analysts, Traders)

• Tech jobs (Programmers, Data analysts, Engineers, IT)

• Teaching jobs (Teachers, Course creators)

• Legal jobs (Lawyers, Legal assistants)

• Health jobs (Doctors, Nurses)

• Sales and Marketing jobs (Sales representatives, Brand managers, Digital marketers)

• Other (please specify)

• I don’t know

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

Subscribe now

American workers’ readiness to use new Generative AI tools is low

The text-to-image platform Midjourney recently celebrated its first birthday while ChatGPT has yet to mark its one-year launch anniversary. Despite these technologies being in their infancy, Americans are expected to adapt swiftly to Generative AI.

In August of this year, Bloomberg heralded, “Generative AI Is Coming to Your Office.” Just one month later Microsoft announced that it would expand its Bing Chat Copilot into a unified Generative AI experience crossing its widely used suite of M365 tools like Word and Excel as well as the Windows operating system. Google Duet — a rebranding of its Generative AI offering — is already available for customers of its paid Google Workspace suite, which includes Gmail, a market leading email client.

The 2023 Generative AI & American Society Survey reveals that just 21% of Americans feel prepared to employ new Generative AI tools, an understandable sentiment considering the newness and fast changing pace of Generative AI technologies. A third of Americans feel unprepared, and 22% are unsure of their preparedness, perhaps because they are not yet familiar enough with Generative AI tools. More than one-fifth (22%) believe they won’t use Generative AI in their job, a somewhat uncertain prospect at this early stage of Generative AI’s emergence.

There was diversity in feeling ready to work with Generative AI among Americans. Younger Americans tended to feel more ready than older Americans; notably just 12% of those 60 and over feel ready to work with Generative AI compared to 27% of those 18 to 29.

Higher income Americans also tend to feel more prepared with 32% of those making $100,000 or more annually feeling ready compared to just 15% of those making less than $30,000. This difference in perception of preparedness could be a misconception; a report from consulting firm McKinsey found that higher-paying jobs are those most likely to be impacted by Generative AI.

Finally, those with less educational attainment tend to feel particularly unready. Only 12% of those with no high school diploma feel ready compared to nearly a third of those with a bachelor’s degree or higher.

When it comes to feeling unprepared there is only slight variation among age groups, income brackets, and education; about a third of Americans feel unready regardless of their demographic characteristics.

Some Americans favor support from AI companies to help those impacted by Generative AI

When asked if AI companies should contribute to a support fund to provide relief to those whose jobs are impacted by Generative AI, just over four in ten Americans (42%) agreed. Another third of Americans (34%) neither agree or disagree with the proposal, while only 12% disagreed.

Public opinion on the idea of an AI support fund shows only minor differences when it comes to agreement across income levels: 42% for Americans with annual incomes exceeding $100,000 compared to 40% for those earning under $30,000. However, when it comes to disagreement, the divergence becomes more noticeable: 18% of the highest earners oppose the policy, more than twice the 8% recorded among the lowest income group.

Agreement was also slightly higher among Americans age 45 and older (44%), more educated individuals (41%), and Black Americans (51%).

Opinions on support from AI companies are shaped by a consideration of both potential benefits and drawbacks. When soliciting responses, the phrasing of the question was designed to encompass these aspects:

Idea: Companies creating AI systems should contribute to a support fund for workers who lose their jobs because of AI.

Benefit: This could help provide a financial safety net for people who lose their jobs. It might also reduce stress as they transition to new job roles.

Tradeoff: This could result in higher prices to use these AI systems. It might also mean fewer AI systems or companies are created.

Share

One-third of Americans favor Government-funded education to help with Generative AI

While there is public backing for government-funded education to prepare for Generative AI, it's not as strong as the support for funding from AI companies. One-third of Americans endorse government-sponsored educational initiatives, compared to a quarter who oppose them. Additionally, nearly one-third of Americans (30%) remain neutral on the proposal.

The level of support for government-funded Generative AI education varies notably across different education levels, income brackets, and racial or ethnic groups. Specifically, 46% of Americans who hold a bachelor's degree or higher support the idea, compared to just 20% of those without a high school diploma. Among income categories, the highest support comes from individuals earning $100,000 or more, at 39%, whereas the support level for those earning less than $30,000 is 30%. Regarding race and ethnicity, individuals who identify as Asian, Native American, or from other racial categories show the highest level of support at 41%.

The situation is nuanced when it comes to opposition to the policy. Interestingly, individuals with higher incomes, who are most likely to support the policy, are also the most likely to oppose it, with 31% disagreeing. This contrasts with the 17% disagreement rate among those earning less than $30,000. Similarly, those with lower educational levels show a moderate level of disagreement at 35%, as do Whites at 30%.

The varying levels of support across income, racial, and educational groups may be influenced by concerns about potential tax increases or cuts to other government programs needed to fund Generative AI education. Typically, individuals with lower education, lower incomes, and non-White ethnicities are more reliant on government benefits,1 which could be affected, while those with more education and higher incomes tend to pay more in taxes.2 When soliciting the level of agreement, the survey question explicitly mentioned these possible trade-offs:

Idea: The government should fund AI education to prepare the workforce.

Benefit: This could reduce unemployment and boost the economy.

Tradeoff: This could result in higher taxes. It could also require cuts to other government programs.

Subscribe now

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. AmeriSpeak is one of the most scientifically rigorous survey options available in the U.S. market. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about AI preparedness and support are below. For more information, please refer to the FAQ page.

Readiness to use Generative AI tools

Think about the next three years. Do you feel ready to use Generative AI tools in your job over that period? Think about tasks like using AI for data analysis, computer programming, writing help, or creative projects.

• Very ready

• Somewhat ready

• Not very ready

• Not ready at all

• I won't use AI in my job

• I don’t know

A total of 13 policy questions were presented in a grid format, grouped into three thematic categories. Each question allowed respondents to choose their level of agreement. The order of the 13 questions presented to respondents was randomized. Each policy question included an idea along with potential benefits and tradeoffs.

Support for funding from AI companies

The following are some ideas for potential changes to how Generative AI systems and companies operate. Each idea has benefits, but there could also be downsides, or "tradeoffs." We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Idea: Companies creating AI systems should contribute to a support fund for workers who lose their jobs because of AI.

Benefit: This could help provide a financial safety net for people who lose their jobs. It might also reduce stress as they transition to new job roles.

Tradeoff: This could result in higher prices to use these AI systems. It might also mean fewer AI systems or companies are created.

•  Strongly agree

• Agree

• Neither agree nor disagree

• Disagree

• Strongly disagree

• Don’t know

Support for Government-funded education programs

The following are some ideas for how governments might handle Generative AI. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Idea: The government should fund AI education to prepare the workforce.

Benefit: This could reduce unemployment and boost the economy.

Tradeoff: This could result in higher taxes. It could also require cuts to other government programs.

•  Strongly agree

• Agree

• Neither agree nor disagree

• Disagree

• Strongly disagree

• Don’t know

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

Subscribe now

Moderately more Americans favor being cautious than favor keeping the U.S. lead in Generative AI

While Generative AI is an emerging technology, the U.S. currently holds a clear advantage in the field. Both ChatGPT, the leading AI language model, and Midjourney, the most popular text-to-image platform, are products of U.S.-based companies. Major tech giants investing in this area, such as Google, Microsoft, and Adobe, are also headquartered in the U.S. Additionally, rising contenders in the space, like Anthropic and Hugging Face, are American-based companies.

New York Senator Chuck Schumer recently discussed America's leading position in Generative AI during a speech at the Center for Strategic and International Studies. Schumer is the sponsor of the SAFE Innovation Framework for Generative AI, legislation aimed at balancing increased regulation with the preservation of the U.S. lead in the field. Senator Schumer noted the following in his remarks:

There are those who fear AI’s immense power and conclude it’s better to turn back, to go no further down this unknown road. We all know it’s not that simple and it’s not that easy. The AI revolution is going to happen with us or without us. If we can promote innovation, make sure it is safe, if America leads the way, the future will be far better, brighter, and safer than if it happens without us.

When asked about the U.S. preserving its global leadership in Generative AI, 36% of Americans lean toward exercising caution, compared to 22% who endorse maintaining our lead.

These cautious views are notable given that Americans were presented with both the potential benefits, such as better living standards and enhanced national security, as well as possible downsides like job loss and AI-generated misinformation. The precise question wording was:

Think about the U.S. keeping its global lead in Generative AI. Benefits could include better living standards, strong economic growth, and improved national security. However, there might also be downsides like job losses, AI-driven cyberattacks, and misinformation created by AI. Given these benefits and downsides, where do you stand?

Men are more likely than women to favor maintaining the U.S. lead, with 28% of male respondents supporting this stance compared to 17% of females. However, a cautious approach garners broader support across genders, with 34% of men and 38% of women favoring caution.

Age also plays a role; the 60 and above age group is most likely to favor a cautious approach, at 42%, while those aged 18-29 are least likely to support either maintaining the lead or being cautious (16% and 25%, respectively). People with higher income levels — those making $60,000 per year and above — also appear more inclined to favor a cautious approach. Specifically, 47% of Americans earning between $60,000 and $100,000 per year, and 41% of those earning above $100,000 per year, advocate for a cautious approach to Generative AI.

Educationally, those with some college education or a bachelor's degree are more likely to both support maintaining the lead and favoring caution, at rates of 20% and 29% for maintaining the lead, and 44% and 40% for caution, respectively. White Americans, more than any other racial or ethnic group, are likely to advocate caution, at 42%.

Americans back global penalties if countries misuse Generative AI

In October 2022, the Biden administration imposed sanctions on China, specifically targeting advanced semiconductors and chip-manufacturing equipment used in AI systems. These measures aim to curb China’s “large-scale surveillance activities enabling human-rights violations and abuses.” Additional AI-focused sanctions were enacted in August 2023.

The 2023 Generative AI & American Society Survey found that a majority of Americans, 54%, support such global penalties, including sanctions, for countries that misuse Generative AI technologies. Only 7% disagree, while 25% neither agree nor disagree.

There are notable demographic differences in attitudes towards imposing global penalties, such as sanctions, on countries that misuse AI technologies.

Age appears to influence opinions: Older age groups (60+) were more likely to agree with such penalties at 61%, whereas the 18-29 age group showed the least agreement at 45%.

Gender also plays a role. Males were more likely to agree with the sanctions, at 59%, compared to females at 51%.

In terms of education, those with a Bachelor's degree or higher were the most likely to agree, at 69%. On the other end, individuals without a high school diploma were the least likely to agree, at 46%.

Among racial and ethnic groups, White Americans showed the highest level of agreement at 60%, while Hispanic and Black respondents were less likely to agree, at 45% and 41% respectively.

Household income also impacts views on this issue. Those in the highest income bracket ($100k+) were the most supportive, with a 66% agreement rate, while those earning less than $30k were least supportive, at 37%.

Overall, these findings suggest that opinions on global penalties for the misuse of AI technologies are shaped by a variety of demographic factors.

Opinions on global penalties included priming respondents to consider both the potential benefits and tradeoffs of such a policy:

Idea: Some countries might misuse AI technology to harm their citizens or threaten democratic values. If they do, these countries should face global penalties, such as sanctions.

Benefit: This could deter nations from abusing AI technology. It might encourage them to follow global democratic principles.

Tradeoff: This could reduce teamwork among nations. Prices could rise from less global trade. It could also spark diplomatic tensions.

Share

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about both the U.S. maintaining its lead and about global penalties are shown below. For more information, please refer to the FAQ page.

U.S. keeping lead question

The survey responses to this question were randomized so that half of respondents saw the scale starting with “Strongly favor keeping our lead” and the other half saw it starting with “Strongly favor being cautious.”

Think about the U.S. keeping its global lead in Generative AI. Benefits could include better living standards, strong economic growth, and improved national security. However, there might also be downsides like job losses, AI-driven cyberattacks, and misinformation created by AI. Given these benefits and downsides, where do you stand?

• Strongly favor keeping our lead

• Somewhat favor keeping our lead

• Neutral

• Somewhat favor being cautious

• Strongly favor being cautious

• I don't know

Global penalties question

A total of 13 policy questions were presented in a grid format, grouped into three thematic categories. Each question allowed respondents to choose their level of agreement. The order of the 13 questions presented to respondents was randomized. Each policy question included an idea along with potential benefits and tradeoffs.

The following are some ideas for how governments might handle Generative AI. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Idea: Some countries might misuse AI technology to harm their citizens or threaten democratic values. If they do, these countries should face global penalties, such as sanctions.

Benefit: This could deter nations from abusing AI technology. It might encourage them to follow global democratic principles.

Tradeoff: This could reduce teamwork among nations. Prices could rise from less global trade. It could also spark diplomatic tensions.

• Strongly agree

• Agree

• Neither agree nor disagree

• Disagree

• Strongly disagree

• Don’t know

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

Subscribe now

Four times as many Americans feel concerned than feel excited about Generative AI

A November 2021 survey from the Pew Research Center found that 37% of adults were more concerned than excited about AI in daily life. However, this survey was conducted before the release of new Generative AI technologies like ChatGPT and Midjourney.

The 2023 Generative AI & American Society Survey revisited the question posed by Pew, but with an emphasis on Generative AI, revealing a comparable level of concern among Americans. The results were strikingly similar: while Pew’s survey indicated that 37% of Americans were more concerned than excited about AI, the Generative AI survey found an analogous 36% of Americans expressing more concern than excitement about Generative AI.

Just 8% were more excited than concerned, while 28% are equally concerned and excited. More than a quarter, 27%, don’t know; likely because not everyone has been exposed fully to new Generative AI technology.

Given the extensive media coverage and discussions surrounding the potential risks of AI, it is not surprising that more Americans have expressed concern than excitement over these technologies. For instance, a widely circulated open letter endorsed by nearly 34,000 signatories—including prominent tech figures like Elon Musk—has raised alarms about the varied threats posed by new Generative AI technologies. These threats range from job displacement and the proliferation of AI-generated misinformation to the severe, existential risk of humanity’s destruction. This sentiment of concern is further echoed in a survey at the well-publicized Yale CEO Summit, where 42% of CEOs concurred that AI harbors the potential to extinguish humanity within the next decade.

Reasons for concern are diverse

To get a sense of why some Americans feel more concerned than excited, the 36% that expressed concern were asked to select the primary reason for concern from among a list of 10 commonly cited fears (there was also an option to select “other”).1 See the project FAQ for why this was a forced choice selection.

While no single fear dominated public concern about AI, the most cited worry focused on the technology's potential to become a potential risk to humanity. This theme has deep roots in American culture, from Isaac Asimov's Three Laws of Robotics, to iconic films like 2001: A Space Odyssey and The Terminator.

More recently fears about Generative AI’s risk to humanity have been articulated by leading figures in the AI industry. A statement from the Center for AI Safety called for prioritizing the existential risks of AI on a global scale, akin to addressing pandemics or nuclear threats. The statement was signed by high-profile experts, including Sam Altman, cofounder of OpenAI, and Geoffrey Hinton, often dubbed as one of the three “Godfathers of AI,” both of whom emphasized the urgent need for action, as reported by multiple media outlets. The upcoming global AI Safety Summit hosted by UK Prime Minister Rishi Sunak is reportedly expected to focus almost entirely on Generative AI’s potential risk to humanity.

Others in the AI community have criticized the focus on existential risk, worrying it diverts attention from more pressing and grounded harms. “AI Causes Real Harm. Let’s Focus on That over the End-of-Humanity Hype,” reads the title of a recent op-ed in Scientific American by AI researchers Emily Bender and Alex Hanna.

While AI’s risk to humanity was the most frequently cited concern, it made up only 16% of the total. Other, more immediate worries took precedence for many respondents including AI misuse (both intentional and accidental), reduced human interaction, job losses, and AI-generated misinformation.2

For example, 15% were primarily concerned about the irresponsible use of AI causing unintended harm, while 14% feared its deliberate misuse. Follow-up research will be needed to explore what harms and misuse most concern Americans.

A further 12% were worried that AI could diminish human interaction and dilute the uniquely human qualities in art, music, and writing and 9% primarily worry about job losses. These apprehensions have taken center stage in the current strikes by writers and actors unions. As production studios intensify their focus on AI-related job openings, a major sticking point in the negotiations is the worry that studios will use Generative AI technologies to automate script-writing tasks.3 Some actors fear that future iterations of Generative AI could eventually replace actors altogether.4

Rounding out concerns are fears about AI-generated misinformation (9%), risks to privacy (6%), bias within AI systems (2%), and the impact of AI on the environment (1%). (I’ve previously written about both AI-generated misinformation and carbon emissions of AI data centers).

Share

Some Americans remain optimistic

While more than a third of Americans are more concerned than excited about AI, 28% are equally concerned and excited and 8% are more excited than concerned. To understand why some Americans feel more excited than concerned, the 8% that expressed excitement were asked to select the primary reason for excitement from among a list of six commonly cited promise areas (there was also an option to select “other”). See the project FAQ for why this was a forced choice selection.

Before discussing these results it’s important to note that not all AI experts fully buy into the belief around AI’s catastrophic potential. Yann LeCun, Meta’s Chief AI Scientist and another of the three “Godfathers of AI,” has stated that fears of AI’s risk to humanity are overblown. He believes that while more immediate concerns like job losses are possible, the long-term benefits of Generative AI outweigh the risks.

In LeCun’s testimony during the Senate’s U.S. Intelligence Committee hearing he stressed the importance of safety, but also access to transformative Generative AI technology, arguing that:

Having access to state of the art AI will be an increasingly important driver of opportunity in the future for individuals, for companies, and for economies as a whole.

Even OpenAI co-founder Sam Altman, who, as discussed earlier, has expressed significant concerns about the negative potential of Generative AI, ultimately believes that Generative AI will be a net positive. In March of 2023, he explained this optimism with noted tech journalist Kara Swisher.

[OpenAI] will be a participant in this technological revolution that I believe will be far greater in terms of impact and benefit than any before…We will be one of several in this moment, and that is going to be really wonderful. This is going to elevate humanity in ways we still can’t fully envision. And our children, our children’s children, are going to be far better off than the best of anyone from this time. And we’re just going to be in a radically improved world. We will live healthier, more interesting, more fulfilling lives; we’ll have material abundance for people…

The opportunity for improved material standard of living and increased fulfilment were echoed by the 8% of Americans expressing optimism. Topping reasons for excitement were various dimensions of Generative AI’s ability to improve productivity. This included saving time by automating mundane tasks (26%), improving productivity at one’s job (18%), and enhancing personal creativity (13%).

Early research indeed suggests that Generative AI has the potential to significantly boost individual productivity and creativity across various professions:

• One study involving 450 college-educated professionals, including marketers, found that using ChatGPT reduced the time needed to complete work-related writing tasks by about 40% and improved the quality of the output.

• Another experiment involving 100 software developers showed that those using GitHub’s Copilot completed their tasks in half the time compared to those coding without it, and a subsequent survey of 2,000 developers indicated improvements in both productivity and job satisfaction.

• An August 2023 preprint presenting the results of a study involving over 200 writers found that narratives created with AI assistance scored higher on creativity metrics than those crafted without AI, suggesting AI enhancement of human creativity.5

• Just a week ago and colleagues released the results of their experiment using ChatGPT-4 at Boston Consulting Group. “Consultants using AI finished 12.2% more tasks on average, completed tasks 25.1% more quickly, and produced 40% higher quality results than those without,” Mollick reported. See the full write-up on Mollick’s substack .

Consumers and workers won’t have to wait long to see these productivity improvements. Just this month Microsoft announced that it would expand its Bing Chat co-pilot into a unified Generative AI experience crossing its widely used suite of M365 tools like Word and Excel as well as the Windows operating system. Google Duet — a rebranding of its Generative AI offering — is already available for customers of its paid Google Workspace suite, which includes Gmail, a market leading email client. Some features from Adobe Firefly’s Generative AI image platform have already been incorporated into Adobe Photoshop, with more likely on the way.

Among those excited about Generative AI, 23% primarily anticipate improved educational tools as a key benefit. Several major EdTech platforms, such as Khan Academy, edX, Coursera, and Pearson, have already announced initiatives to integrate Generative AI language technologies into their services.6 Mobile applications are also getting in on the action; an early entry is the Hello History app, which uses an AI language model to let users engage in conversations with historical figures.

Improving accessibility to digital tools for people with disabilities (10%) and providing unique entertainment content (10%) rounded out reasons for excitement.

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about AI excitement and concern are shown below. For more information, please refer to the FAQ page.

Concern or excitement question

The concern question was randomized so that about half of respondents saw “More excited than concerned” as the first option and the other half saw “More concerned than excited” first.

As new Generative AI technologies are becoming more common in our everyday lives, how does this make you feel?

• More excited than concerned

• More concerned than excited

• Equally concerned and excited

• I don’t know

Reasons for excitement question

This question was shown to those respondents who selected “More excited than concerned.” Response options were shown to respondents in a random order.

What is the main reason that you are more excited than concerned about the increased use of Generative AI in daily life?

• It could improve my productivity at my job (for example, writing emails, creating presentations, doing research).

• It could enhance my personal creativity (for example, in music, writing, or visual arts).

• It could provide unique entertainment content (for example, stories, music, visual art).

• It could increase accessibility of digital tools for people with disabilities.

• It could serve as a valuable educational tool, assisting in learning new subjects or concepts.

• It could save me time by automating mundane tasks.

• Other (please specify)

Reasons for concern question

This question was shown to those respondents who selected “More concerned than excited.” Response options were shown to respondents in a random order.

What is the main reason that you are more concerned than excited about the increased use of Generative AI in daily life?

• Job losses due to AI.

• Infringements on digital privacy, increased risk of surveillance or hacking.

• Reduced human interaction or loss of uniquely human aspects in art, music, and writing.

• AI becoming too powerful and becoming a potential risk to humanity.

• The irresponsible use of AI leading to unintended harm.

• AI being misused to cause intentional harm.

• The potential for bias within AI systems.

• The energy and resource consumption of AI systems and their potential impact on the environment.

• The spread of misinformation or false content created by AI.

• The lack of adequate oversight and regulation of AI.

• Other (please specify)

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

Subscribe now

American interest in Generative AI tools has soared, but not everyone uses them regularly

Interest in Generative AI has soared in 2023 as tools like ChatGPT and Midjourney were thrust into the public consciousness. While Google and Microsoft publicly contended to release competing AI language tools in Bard and Bing Chat, respectively, ChatGPT has led the charge (as evidence, The New York Times now has a technology subsection specifically dedicated to ChatGPT). There is a long-tail of lesser-known AI language tools like search engines You.com and Perplexity, coding plugin GitHub Co-pilot, and a bevy of new tools claiming to leverage AI for writing help. Among AI text-to-image tools Midjourney is the clear winner in public interest thus far, beating out other common tools like Adobe Firefly and DALL-E 2 from OpenAI. (However, OpenAI just announced DALL-E 3, which will be integrated into ChatGPT).

Subscribe now

While there is some information on the use of specific Generative AI tools — ChatGPT, for instance, attracts approximately 18 million unique monthly visitors in the United States according to SimilarWeb1 — comprehensive data on self-reported usage across all types of AI language and text-to-image tools has been unavailable until now.

The Generative AI & American Society Survey results indicate that one in five American adults use AI language models at least monthly, while AI text-to-image tools see only half as much use, at 10%. This amounts to about 52 million monthly active users of AI language tools and 26 million monthly active users of AI text-to-language tools.2 Half of American adults have not yet interacted with these Generative AI tools despite being aware of them.

In spite of their significant media coverage, awareness of these technologies is not universal. One-fifth of Americans haven't heard of AI language models, and a full one-third are unfamiliar with AI text-to-image tools.

This is likely our last look at Generative AI while standalone technologies are the primary mode of usage, before they proliferate fully into consumer software products. Soon consumers may begin to leverage Generative AI tools without even being fully aware they’re doing so. Just last week Microsoft announced that it would expand its Bing Chat co-pilot into a unified Generative AI experience crossing its widely used suite of M365 tools like Word and Excel as well as the Windows operating system. Google Duet — a rebranding of its Generative AI offering — is already available for customers of its paid Google Workspace suite, which includes Gmail, a market leading email client. Some features from Adobe Firefly’s Generative AI image platform have already been incorporated into Adobe Photoshop, with more likely on the way.

Lower income Americans are more likely to have not heard of AI language tools

While the familiarity with AI text-to-image tools is consistent across income brackets, ranging between 30% to 35%, the same cannot be said for AI language tools. Lower-income Americans — those earning less than $30,000 annually — are about twice as likely to have never heard of language-based AI tools like ChatGPT when compared to those with incomes of $100,000 or more.

This same pattern holds true across educational attainment; 32% of those that didn’t finish high school have never heard of Generative AI tools compared to just 13% for those with a bachelors degree or higher.

Perhaps surprisingly given the figures above, those earning less than $30,000 are the most likely to use AI language models everyday at 8%, compared to 3% for those making more than $100,000. (However, note that while this difference is statistically significant, the sample size for this detailed demographic breakdown is quite small, and therefore, may be less accurate).

Share

White Americans use AI language tools less frequently

In terms of regular use of AI language tools, White Americans trail other racial and ethnic groups. While 13% of White Americans use these tools at least once a month, the rates are 29% for Black Americans, 24% for Hispanics, and 36% for Asians and other ethnicities. Moreover, White Americans are the group most likely to be aware of but not use these tools, with 55% falling into this category.

Older Americans are also less likely to have heard of new Generative AI technologies. Among those aged 60 and over, 27% have never heard of AI language tools, while a larger 41% remain unaware of AI text-to-image tools. A mere 14% haven't heard of AI language tools, although a somewhat larger percentage — 26% — haven't heard of AI text-to-image tools. Those aged 18 to 29 have similar rates of knowledge.

Subscribe now

How were these figures determined?

Results from the 2023 Generative AI & American Society Survey came from the National Opinion Research Center’s probability-based AmeriSpeak panel of adults ages 18 and over. The sample size was 1,147 and responses were weighted to ensure national representation. The exact questions posed to respondents about both AI language tools and AI text-to-image tools are detailed below. For more information, please refer to the FAQ page.

AI language tools question

Think about the past three months. In an average month, how often have you used language-based AI systems? These are systems that can chat like a human and perform other tasks such as writing computer code. Examples include ChatGPT by OpenAI, Bard from Google, and Bing Chat from Microsoft.

• Every day

• Several times a week

• Once a week

• 1-3 times a month

• I have tried them but did not continue using them.

• I'm aware of them, but I have not used them.

• I've never heard of them.

AI text-to-image tools question

Think about the past three months. In an average month, how often have you used AI tools that make images from a text description entered by a human? Examples include Midjourney, Adobe Firefly, and DALL-E 2 by OpenAI.

• Every day

• Several times a week

• Once a week

• 1-3 times a month

• I have tried them but did not continue using them.

• I'm aware of them, but I have not used them.

• I've never heard of them.

If you have additional questions, comments, or suggestions please do leave a comment below or email me at james@96layers.ai. To help advance the understanding of public attitudes about Generative AI I’m making all raw data behind the 2023 Generative AI & American Society Survey available free of charge. Please email me if you’re interested.

He chuckled. "Old habits die hard," he mumbled to himself.

Mark had been thinking a lot about switching career lanes to something more creative, perhaps writing. But the survey didn't bode well for that idea either. Apparently, 21% of respondents believed that creative professions like artists and writers would be negatively impacted by Generative AI.

His phone buzzed. A message from his friend Emily, a game designer, flashed on the screen.

"Hey, you won't believe what happened at work today! One of our new interns is actually a Generative AI model. Can you believe it?"

Mark chuckled and texted back, "Actually, I can. More people seem to think your job will benefit from AI than be harmed. Just saw a survey."

While pondering his own career choices, Mark also wondered about his sister, who was a teacher. The survey indicated a nearly split opinion on jobs like hers—5-7% believed teaching would benefit from AI, while a similar percentage felt it would suffer.

His thoughts were interrupted by a news notification.

"Ongoing strikes by writers and actors against AI usage continue to draw crowds," it read.

Mark was puzzled. Only 5% in the survey thought performing arts would be most harmed. Was this a case of the minority shouting the loudest?

He sighed, closed his laptop, and looked out the window. The sun was setting, casting long shadows over the city, much like the uncertainties AI was casting over various careers.

Mark knew choices had to be made, and they had to be made soon. AI wasn't coming; it was already here. And whether it would be a tool or a competitor was a decision that couldn't be made by statistics alone.

"Uncertainty might be the only certainty in this AI-driven world," he thought as he reopened his laptop to study up on Generative AI. "Best be ready for whatever comes next."

Subscribe now

The following questions are about Generative Artificial Intelligence (AI). Generative AI is a new kind of technology. It can produce things like text, computer code, images, music, voices, and videos. It does this by getting some guidance from humans. For example, you might have heard of AI systems like ChatGPT. It can chat like a human. Another tool, called Midjourney, can make images based on text descriptions entered by a person.

These Generative AI tools learn from a lot of data. This data could be books, articles, pictures, or songs. From this data, the AI learns patterns. After learning, they can create new content. This content is similar to the data they learned from. For example, if the AI learned from many books, it could write a story. If it learned from images, it can create new images. The same goes for music, computer code, and simple videos.

These questions aren’t about all AI. They don’t cover things like self-driving cars or facial recognition. They’re only about new Generative AI that can create human-like content. Keep that in mind when you answer the questions.

Question 1

Think about the past three months. In an average month, how often have you used language-based AI systems? These are systems that can chat like a human and perform other tasks such as writing computer code. Examples include ChatGPT by OpenAI, Bard from Google, and Bing Chat from Microsoft.

• Every day

• Several times a week

• Once a week

• 1-3 times a month

• I have tried them but did not continue using them.

• I'm aware of them, but I have not used them.

• I've never heard of them.

Question 2

Think about the past three months. In an average month, how often have you used AI tools that make images from a text description entered by a human? Examples include Midjourney, Adobe Firefly, and DALL-E 2 by OpenAI.

• Every day

• Several times a week

• Once a week

• 1-3 times a month

• I have tried them but did not continue using them.

• I'm aware of them, but I have not used them.

• I've never heard of them.

Question 3

As new Generative AI technologies are becoming more common in our everyday lives, how does this make you feel?

• More excited than concerned

• More concerned than excited

• Equally concerned and excited

• I don’t know

Question 4 (if excited)

What is the main reason that you are more excited than concerned about the increased use of Generative AI in daily life?

• It could improve my productivity at my job (for example, writing emails, creating presentations, doing research).

• It could enhance my personal creativity (for example, in music, writing, or visual arts).

• It could provide unique entertainment content (for example, stories, music, visual art).

• It could increase accessibility of digital tools for people with disabilities.

• It could serve as a valuable educational tool, assisting in learning new subjects or concepts.

• It could save me time by automating mundane tasks.

• Other (please specify)

Question 4 (if concerned)

What is the main reason that you are more concerned than excited about the increased use of Generative AI in daily life?

• Job losses due to AI.

• Infringements on digital privacy, increased risk of surveillance or hacking.

• Reduced human interaction or loss of uniquely human aspects in art, music, and writing.

• AI becoming too powerful and becoming a potential risk to humanity.

• The irresponsible use of AI leading to unintended harm.

• AI being misused to cause intentional harm.

• The potential for bias within AI systems.

• The energy and resource consumption of AI systems and their potential impact on the environment.

• The spread of misinformation or false content created by AI.

• The lack of adequate oversight and regulation of AI.

• Other (please specify)

Question 5

[Question 5 was presented in grid format with the following response options: “Strongly agree,” “Agree,” “Neither agree nor disagree,” “Disagree,” “Strongly Disagree,” “I don’t know.”]

The following are some ideas to make AI systems more open and responsible. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Question 5a.

Idea: AI systems should clearly explain how they make decisions.

Benefit: Clear AI explanations could help users make smarter choices.

Tradeoff: This could make AI systems harder to build so they might cost more to use.

Question 5b.

Idea: Companies should tell us how their AI models work and what data they use.

Benefit: This approach could foster trust and allow users to pick the AI systems they feel comfortable using.

Tradeoff: This could give away AI company secrets and let AI creators in other countries get ahead. In addition, knowing they have to disclose details might make companies less eager to innovate.

Question 5c.

Idea: AI companies should tell users when they're talking to an AI, not a person.

Benefit: Knowing they're talking to an AI could help users set clear expectations about the interaction.

Tradeoff: This could break the feel of a real chat and make AI systems less enjoyable or easy to use.

Question 5d.

Idea: Content created by AI should have a clear label saying it's AI-generated. AI-generated content includes images, videos, stories, and music.

Benefit: This could enable users to understand content origins and make informed decisions.

Tradeoff: People might engage with content less if they know it's made by AI even if the content is useful or entertaining. That's because it could seem less genuine or appealing.

Question 6

[Question 6 was presented in grid format with the following response options: “Strongly agree,” “Agree,” “Neither agree nor disagree,” “Disagree,” “Strongly Disagree,” “I don’t know.”]

The following are some ideas for potential changes to how Generative AI systems and companies operate. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Question 6a.

Idea: AI systems should undergo thorough checks for bias and fairness before their launch.

Benefit: If AI is fair, it could mean everyone gets an equal shot, especially when big decisions are made.

Tradeoff: This could delay the release of the AI system or make the system less creative.

Question 6b.

Idea: AI systems should be designed to prioritize user privacy and data security.

Benefit: This could help keep personal data out of the wrong hands.

Tradeoff: This might make the AI less useful or harder to use.

Question 6c.

Idea: Companies creating AI systems should contribute to a support fund for workers who lose their jobs because of AI.

Benefit: This could help provide a financial safety net for people who lose their jobs. It might also reduce stress as they transition to new job roles.

Tradeoff: This could result in higher prices to use these AI systems. It might also mean fewer AI systems or companies are created.

Question 6d.

Idea: Artists, authors, and musicians should be paid if AI systems use their works to learn. Or they should have the choice to opt-out.

Benefit: Creators could get recognized and financially rewarded for their work.

Tradeoff: This could limit the diversity of AI training data, affecting the quality of AI content. It might also increase the cost of using AI systems.

Question 6e.

Idea: AI tools generating content should have built-in safeguards to avoid creating harmful or offensive content.

Benefit: This could result in content that is more respectful and less likely to cause emotional harm.

Tradeoff: This could mean that companies define what is considered “harmful” or “offensive”. This might lead to censorship issues.

Question 7

[Question 7 was presented in grid format with the following response options: “Strongly agree,” “Agree,” “Neither agree nor disagree,” “Disagree,” “Strongly Disagree,” “I don’t know.”]

The following are some ideas for how governments might handle Generative AI. Each idea has benefits, but there could also be downsides, or "tradeoffs". We've included potential benefits and tradeoffs for context, but there might be more that we haven’t mentioned.

Indicate how much you agree or disagree with each idea. Keep in mind, your opinion is only about the idea itself. The benefit and tradeoff might help guide your opinion.

Question 7a.

Idea: Regulations should stop a few technology companies from controlling the entire AI market.

Benefit: This could boost competition. It could allow for more diverse AI solutions. It could also prevent one company from getting too powerful.

Tradeoff: This could slow down the pace of AI development and innovation.

Question 7b.

Idea: Some countries might misuse AI technology to harm their citizens or threaten democratic values. If they do, these countries should face global penalties, such as sanctions.

Benefit: This could deter nations from abusing AI technology. It might encourage them to follow global democratic principles.

Tradeoff: This could reduce teamwork among nations. Prices could rise from less global trade. It could also spark diplomatic tensions.

Question 7c.

Idea: The government shouldn't let AI assist in running crucial systems. These include power grids, transit, telephone lines, or the internet.

Benefit: By excluding AI, we might reduce risks linked to AI errors or cyber attacks in these systems.

Tradeoff: Lack of AI in these systems might reduce automation and increase user costs. For example, your power bill could be higher than if AI were used.

Question 7d.

Idea: The government should fund AI education to prepare the workforce.

Benefit: This could reduce unemployment and boost the economy.

Tradeoff: This could result in higher taxes. It could also require cuts to other government programs.

Question 8

Think about the U.S. keeping its global lead in Generative AI. Benefits could include better living standards, strong economic growth, and improved national security. However, there might also be downsides like job losses, AI-driven cyberattacks, and misinformation created by AI. Given these benefits and downsides, where do you stand?

• Strongly favor keeping our lead

• Somewhat favor keeping our lead

• Neutral

• Somewhat favor being cautious

• Strongly favor being cautious

• I don't know

Question 9

Think about the next three years. Which job area do you believe will benefit the most from Generative AI? For each job area we've given some example jobs, but there might be other jobs in the job area.

• Creative jobs (Artists, Writers, Designers)

• Media jobs (Game designers, Game programmers, 3D artists)

• Performing arts jobs (Movie and TV actors)

• Finance jobs (Accountants, Stock analysts, Traders)

• Tech jobs (Programmers, Data analysts, Engineers, IT)

• Teaching jobs (Teachers, Course creators)

• Legal jobs (Lawyers, Legal assistants)

• Health jobs (Doctors, Nurses)

• Sales and Marketing jobs (Sales representatives, Brand managers, Digital marketers)

• Other (please specify)

• I don’t know

Question 10

Think about the next three years. Which job area do you believe will be most hurt by Generative AI? For each job area we've given some example jobs, but there might be other jobs in the job area.

• Creative jobs (Artists, Writers, Designers)

• Media jobs (Game designers, Game programmers, 3D artists)

• Performing arts jobs (Movie and TV actors)

• Finance jobs (Accountants, Stock analysts, Traders)

• Tech jobs (Programmers, Data analysts, Engineers, IT)

• Teaching jobs (Teachers, Course creators)

• Legal jobs (Lawyers, Legal assistants)

• Health jobs (Doctors, Nurses)

• Sales and Marketing jobs (Sales representatives, Brand managers, Digital marketers)

• Other (please specify)

• I don’t know

Question 11

Think about the next three years. Do you feel ready to use Generative AI tools in your job over that period? Think about tasks like using AI for data analysis, computer programming, writing help, or creative projects.

• Very ready

• Somewhat ready

• Not very ready

• Not ready at all

• I won't use AI in my job

• I don’t know

The purpose of this survey was to poll American attitudes specifically about the new class of artificial intelligence known as Generative AI. Such technologies include ChatGPT, Bing Chat, Midjourney, and Adobe Firefly. To reinforce this concept, many of the questions specifically mention Generative AI technologies. Additionally, there was a preamble at the beginning of the survey which defined its scope. All survey respondents were shown this preamble before continuing on to the questions. The preamble read as follows:

The following questions are about Generative Artificial Intelligence (AI). Generative AI is a new kind of technology. It can produce things like text, computer code, images, music, voices, and videos. It does this by getting some guidance from humans. For example, you might have heard of AI systems like ChatGPT. It can chat like a human. Another tool, called Midjourney, can make images based on text descriptions entered by a person.

These Generative AI tools learn from a lot of data. This data could be books, articles, pictures, or songs. From this data, the AI learns patterns. After learning, they can create new content. This content is similar to the data they learned from. For example, if the AI learned from many books, it could write a story. If it learned from images, it can create new images. The same goes for music, computer code, and simple videos.

These questions aren’t about all AI. They don’t cover things like self-driving cars or facial recognition. They’re only about new Generative AI that can create human-like content. Keep that in mind when you answer the questions.

What questions were asked?

A friendly version of the survey introduction and questions can be found here.

Is this survey really nationally representative of all Americans?

Yes. The 2023 Generative AI and American Society Survey was administered by the National Opinion Research Center (NORC) at the University of Chicago using their AmeriSpeak panel and results are weighted to be nationally representative. It is one of the highest-quality survey options available in the United States. NORC surveys are routinely cited in the media, with over 1,400 total mentions. A technical overview of the AmeriSpeak panel can be found here. For more details see the FAQ question below, “Did this survey use probability sampling?”

Here is NORC’s summary of the AmeriSpeak panel:

Since its founding by NORC at the University of Chicago in 2015, AmeriSpeak has produced more than 900 surveys, been cited by dozens of media outlets and become the primary survey partner of the nation's preeminent news service, The Associated Press. Our sampling captures a true picture of America, providing better representation than other panels for hard-to-reach populations, including low-income households, less educated persons, young adults, rural households, persons who are less interested in the news, and social and political conservatives. AmeriSpeak is the most scientifically rigorous multi-client panel available in the U.S. market.

What are the survey details?

The survey was administered by the National Opinion Research Center (NORC) at the University of Chicago using the AmeriSpeak Omnibus panel. A technical overview of the AmeriSpeak panel can be found here.

Additional details:

• The dates of the survey were 24 August 2023 (a Thursday) to 28 August 2023 (a Monday).

• The survey was administered to U.S. adults ages 18 and over. In this survey the youngest respondent was 18 and the oldest was 95.

• In terms of the device used to respond to the survey, 30% used a desktop, 5% used a touchtone phone (those without internet), and 65% used a mobile device like a smart phone.

• The survey was weighted to ensure it is nationally representative.

• The margin of error is +/-3.87.

• The sample size was 1,147.

• Questions were randomized to try and minimize certain types of response bias. See specific FAQ below for randomization details.

What were the response rates to the survey?

Note that the American Association for Public Opinion Research (AAPOR) recommends the following approach to interpreting survey response rates:

Consumers of survey results should treat all response rates with skepticism, since these rates do not necessarily differentiate reliably between accurate and inaccurate data. Instead consumers should pay attention to other indicators of quality that are included in reports and on websites, such as insignificant levels of bias, low levels of missing data, and conformity with other research findings.

Nonetheless AAPOR also recommends reporting response rate as a best practice. A detailed breakdown of the 2023 Generative AI and American Society Survey response rates can be found below:

1. Fielding Dates: August 24th to August 28th

   • The survey was open for responses during these dates.

2. Released Sample: 6,667

   • The survey was sent out to a sample of 6,667 people.

3. Completes: 1,147

   • 1,147 people completed the survey in full.

4. Recruitment Rate: 21.1%

   • This represents the portion of people who started the survey out of the initial sample. Calculated as (People who started the survey / Released Sample) * 100. This helps to understand the engagement level.

5. Retention Rate: 78.7%

   • This shows the portion of people who completed the survey once they started. Calculated as (Completes / People who started the survey) * 100. This can indicate how straightforward or challenging the survey was to complete.

6. Completion Rate: 17.2%

   • This indicates the portion of people who completed the survey out of the initial sample. Calculated as (Completes / Released Sample) * 100. This provides a broad view of the survey’s effectiveness.

7. AAPOR3 Response Rate: 2.9%

   • AAPOR stands for American Association for Public Opinion Research. The AAPOR3 rate is a specific formula for calculating the response rate.

8. DEFF: 1.79

   • This typically refers to the Design Effect, which measures how much the survey design deviates from a simple random sample. It's often used to adjust the Margin of Error (MOE).

9. MOE: +/- 3.87

   • The Margin of Error means that the survey results can deviate by this percentage in either direction. For example, if 50% of respondents say "Yes" to a question, the real answer, if every person in the population were asked, would likely fall between 46.13% and 53.87%.

Why does the sample size vary between questions?

Respondents have the option to skip any question or to refuse to answer any question (these are two distinct options). The number of respondents that skipped a question ranged from 2 to 27 depending on the specific question, with an average of 16 respondents skipping per question, or just over 1% of the total sample size. Although respondents also had the option to refuse to answer questions, only one respondent exercised this option for a single question (the question about artist compensation).

Additionally, when breaking down specific response options or demographic groups the sample size will vary.

Did this survey use probability-based sampling?

Yes.

Surveying an entire population can be costly, time-consuming, and arduous. Sampling a smaller group from within that population is usually a more realistic option. The purpose of a survey, therefore, is to extrapolate the views of a small group to the views of an entire population. (A survey that includes all members of a population is called a census). There are two broad methodologies to select participants for a survey:

• Probability-based sampling. Probability-based sampling uses a so-called “sampling frame,” a list of the entire population that theoretically could be included in the survey sample. Survey participants are then selected at random from the list. This allows for a direct mathematical calculation to extrapolate the responses of the survey participants to the views of the entire population of interest (called “weighting” the survey). The National Opinion Research Center (NORC) AmeriSpeak panel is a probability-based survey. For technical details see this white paper on AmeriSpeak recruitment.

• Non-probability sampling. Surveys using non-probability sampling are typically cheaper, faster, and easier to conduct than those based on probability sampling. A common type of survey based on non-probability sampling is an online opt-in connivence sample. For instance, a questionnaire a company may send out to all customers that purchased a product within the past month. There are methods to weight surveys based on non-probability sampling, but unlike probability-based sampling there is no universal mathematical approach that can be used in all cases.

Both probability and non-probability sampling methods have their respective merits and appropriate applications. However, when the objective is to acquire results that are representative of the entire population, such as in the Generative AI & American Society Survey, probability-based sampling is generally favored. This method ensures every individual within the population has an equal chance to be included, hence it mitigates the risk of underrepresenting certain groups that is often inherent in non-probability convenience samples.

For an accessible overview of how probability and non-probability sampling works with specific, easy-to-follow examples, see “Sampling: Why and How of it?” For an overview of probability and non-probability sampling considerations see a relatively friendly overview from the American Association of Public Opinion Research, "Summary Report of the AAPOR Task Force on Non-Probability Sampling" which appeared in the Journal of Survey Statistics and Methodology in 2013.

Numerous studies have compared the accuracy of probability and non-probability sampling methods. A full survey of that literature is beyond the scope of this article, but I have provided a short list of articles for those who are interested:

• In September of 2023 the Pew Research Center compared three probability-based surveys with three non-probability surveys on a benchmark of 28 variables. The study found that, “Opt-in samples are about half as accurate as probability-based panels.” This was a follow-up to their 2016 analysis on the same topic.

• A 2016 discussion paper called “Establishing the accuracy of online panels for survey research” compared 18 opt-in surveys and three probability-based surveys against a benchmark government registry of all 16 million Dutch residents (the so-called Municipal Basic Administration, or MBA). The paper summarized its findings as follows:

  The nonprobability samples yielded less accurate estimates of proportions and notably different relations between variables than did the probability samples, and these differences were not eliminated by weighting. These findings reinforce the value of scientific, random sampling to permit generalizing research findings to a larger population.

• A 2020 study compared official data collected by the government of Finland with two non-probability river samples, one with a group of users recruited from Facebook and another recruited via a link on the website of Finish broadcasting company YLE. All three groups were asked about their experience with cyberhate and cyberharrasment. Official data show that just 5% of youth have been exposed to cyberharrasment, but the Facebook survey pegged the figure at 36% (7 times higher) while the YLE survey resulted in an estimate of 30% (6 times higher). Statistical adjustments to weight the river samples did not impact the Facebook estimate, but reduced the YLE estimate to 21% (still 4 times higher). Figures for cyberhate were similarly skewed. The study emphasized that non-probability samples have their place, noting that, “[N]on‐probability online surveys do not replace probability surveys, but augment the researcher's toolkit with new digital practices, such as exploratory studies of small and emerging non‐demographic subpopulations.”

• An analysis in 2021 by Peter K. Enns and Jake Rothschild found that non-probability samples can perform well in certain scenarios, actually out-performing probability-based samples in 2020 election polls. However, the analysis was done at the aggregate level and no demographic segmentation was analyzed.

• A 2022 study compared attitudes of water quality using two opt-in surveys — one from survey provider Qualtrics and oen Amazon’s Mechanical Turk service — against a probability sample. The study found, “The samples differ in some key demographics, but measured attitudes are strikingly similar…Overall, the evidence is mixed, with some key differences but many similarities across samples.”

The 2016 discussion paper cited above concisely summarized when the results of probability and non-probability samples are expected to differ:

In theory, opt-in samples may sometimes yield results that are as accurate as probability samples, assuming that the factors that explain a population member’s presence or absence in the sample are uncorrelated with variables measured in a study and the magnitudes of associations between pairs of variables.

What is Total Survey Error and how did the survey minimize it?

Every survey inherently has imperfections and is susceptible to a spectrum of errors collectively known as Total Survey Error (TSE). While the emphasis of probability-based sampling is fundamentally on minimizing sampling error, it is crucial to recognize that several other types of error can affect a survey’s accuracy. Examples of errors within the TSE framework include:

1. Construct Validity Error: Does a survey question accurately measure the intended concept?

2. Measurement Error: Are respondents providing accurate answers to the questions?

3. Processing Error: Is the data accurately analyzed and represented in the survey results?

4. Coverage Error: Does the sampling frame encompass the entire target population?

5. Sampling Error: Is the sample an accurate representation of the target population?

6. Nonresponse Error: Are the results biased due to the systematic absence of responses from certain segments of the target population?

All of these errors can impact the reliability and validity of survey results. In administering the 2023 Generative AI and American Society Survey I relied on NORC to help minimize processing error, coverage error, and sampling error. I attempted to minimize construct validity error and measurement error via careful crafting of the question wording, ordering, and randomization. Details are provided in the FAQs that follow. Any nonresponse error remains, see the “What were the response rates to the survey?” FAQ for details.

For a good history and overview of Total Survey Error, see “Total Survey Error: Past, Present, and Future.”

Who wrote the survey questions?

The survey was written by James McCammon, a really cool, smart guy and author of 96layers. Questions went through multiple reviews with the National Opinion Research Center as well as a single round of reviews with three volunteer reviewers from the AI community that I cold emailed (reviewers were affiliated with the Centre for the Governance of AI, UC Berkeley AI Policy Hub, and HuggingFace). Many of the FAQ below answer specific questions about the design of the survey questions.

Can I see the raw data files?

Yes. I am making all raw data from the survey available free of charge. The only condition is that if the data is reused, attribution is granted under a Open Data Commons Attribution License (ODC-By). The goal of sharing the data is to advance understanding of American public opinion on AI so that researchers, policy makers, and other interested parties have full insight to the public’s beliefs and attitudes about new Generative AI technologies.

Will there be future versions of this survey?

Yes. As funding allows I will continue to improve and iterate on the survey questions and work with NORC to administer future versions.

Can I contribute to questions in the next iteration of the survey?

Yes! No survey is perfect. If you have ideas about how the survey can be improved leave a comment below or email me at james@96layers.ai. If possible I will try to incorporate your suggestions.

Were questions and response options randomized?

Yes. It is common practice to randomize questions and response options to avoid various kinds of survey bias.

Here is Harvard Professor Stefanie Stantcheva’s advice on randomization from her wonderful paper “How to Run Surveys.”

[T]he order in which response options are provided may not be neutral. Respondents may tend to pick the last answer (a “recency effect” most often encountered in phone or face-to-face surveys) or the first answer (a “primacy effect”). In addition to the advice of avoiding long lists of options and using forced-choice instead of select-all-that-apply, it makes sense to randomize the order of answer options for questions that do not have a natural ordering or where the ordering can be inverted.

Not all questions in the survey were randomized, but many were:

• Question 3 was about excitement or concern over Generative AI. In Question 3 the order of the response options “More excited than concerned” and “More concerned than excited” were randomized so that each appeared first about half the time.

• Question 4 was about the reason for excitement or concern. In Question 4 all response options were randomized for both versions.

• Questions 5, 6, and 7 were about various policy proposals. The order that these questions were presented to respondents was randomized. Additionally, the order of each grid item within the questions was also randomized. All three questions were split across two screens for readability.

• Question 8 was about the U.S. maintaining its lead in Generative AI. In Question 8 the scale was randomized so that half the time it started with “Strongly favor keeping our lead” and half the time it was reversed and started with “Strongly favor being cautious.”

• Questions 9 and 10 were about job areas that would benefit or be hurt by Generative AI. The order that these questions were presented to respondents was randomized. Additionally, within each question the response options were randomized.

Why is the question wording so simple?

Throughout the survey you may notice that the wording of questions is intentionally simple, with short sentences and easy-to-understand language. The question phrasing can be so simple as to seem almost silly at times, but this was very intentional. Survey questions meant to be representative of all Americans need to be accessible to all Americans.

Keep in mind that because the NORC AmeriSpeak survey uses a probability-based, nationally representative panel, respondents have a variety of educational backgrounds. U.S. literacy rates are modest, so surveys need to be constructed to reflect this. Moreover, survey questions have to be simple enough so that the survey can be successfully administered over an automated landline phone system to account for respondents that do not have internet access (about 7% of Americans still don’t use the internet). In the 2023 Generative AI and American Society Survey 5% of respondents completed the survey by phone.

After the initial draft of the survey was completed a tremendous amount of time was put into simplifying the question wording to be as accessible as possible. As an example, here is the original and final version of Question 7b.

Question 7b original:

Countries that misuse AI technology to harm their citizens or threaten democratic values should face international consequences, such as sanctions. However, this approach could lead to less international teamwork on global problems. It might also cause negative economic impacts due to reduced trade between countries. Lastly, it could create diplomatic tensions among nations.

Question 7b final:

Idea: Some countries might misuse AI technology to harm their citizens or threaten democratic values. If they do, these countries should face global penalties, such as sanctions.

Benefit: This could deter nations from abusing AI technology. It might encourage them to follow global democratic principles.

Tradeoff: This could reduce teamwork among nations. Prices could rise from less global trade. It could also spark diplomatic tensions.

The tool Readable was used to assess the expected reach and Flesch-Kincaid Grade Level on a question-by-question basis. The goal was to make each survey question at or below a 9th grade level. According to Readable the overall survey has 100% with a Flesch-Kincaid Grade Level of 7.0.

How was the order of the questions chosen?

Here is Harvard Professor Stefanie Stantcheva’s describe’s question ordering in her wonderful paper “How to Run Surveys.”

In general, the ordering of the questionnaire needs to be guided by three (sometimes conflicting) concerns: i) Respondents are often more engaged and less tired earlier in the survey; ii) at the same time, questions that come earlier can influence responses to subsequent questions through the channels explained in Section 5.2; and iii) respondents form an opinion about your survey in the first few questions, and capturing their interest is critical. Therefore, balancing these three concerns, you must tailor the ordering to your specific question and setting. It is thus difficult to give general advice.

• Questions 1 and 2 were about usage of new AI tools. These questions were asked first to help set the stage for the purpose of the survey.

• Questions 3 and 4 were about excitement and concern. These questions were presented next to continue to build on usage before the policy-focused questions were presented. Additionally, reasons for excitement and concern about Generative AI are of primary interest to a wide audience and so I wanted to place these close to the beginning of the survey.

• Questions 5, 6, and 7 were a set of policy questions. These questions are more cognitively demanding, but still important. It made sense to have these in the middle of the survey after Questions 1-4 had oriented respondents. They could have potentially gone last, but I didn’t want to put the most cognitively demanding questions at the end.

• Question 8 is about the U.S. maintaining its lead. This was a stand-alone policy question that naturally fit after the larger set of related policy questions.

• Questions 9 and 10 were about job impact areas. These could’ve been placed before the policy questions, but again I judged Questions 9 and 10 as less cognitively demanding so chose to place them at the end.

• Question 11 is a natural follow-on to Questions 9 and 10 and so was therefore placed last.

What did the survey questions look like for respondents?

Below is an example of a grid policy question that was presented to respondents.

Why is there white space before some survey responses?

Survey questions have a space before responses options. For example, Question 8 has a space between the set of five main options and “I don’t know.” The reason for this is to match the visual midpoint (“Neutral”) and the contextual midpoint (also “Neutral”). Including “I don’t know” along with the five main response options creates a discontinuity in this respect as “Neutral” remains the contextual midpoint, but is no longer the visual midpoint. Aligning the two reduces cognitive burden and makes the survey friendlier to respondents.

Think about the U.S. keeping its global lead in Generative AI. Benefits could include better living standards, strong economic growth, and improved national security. However, there might also be downsides like job losses, AI-driven cyberattacks, and misinformation created by AI. Given these benefits and downsides, where do you stand?

• Strongly favor keeping our lead

• Somewhat favor keeping our lead

• Neutral

• Somewhat favor being cautious

• Strongly favor being cautious

• I don't know

Contrast Question 8 with the simple survey below from Google. Here the contextual midpoint is “Neutral” but the visual midpoint is “Somewhat Likely.”

Once a space was added before “I don’t know” for Question 8, the same strategy was applied to other questions to maintain continuity.

Even cases without a clear contextual midpoint separating sets of response options improves flow. Consider Question 1, which has all usage options related to frequency in a grouping and a second set of options relating to nonuse. The two sets are separated by a space.

Think about the past three months. In an average month, how often have you used language-based AI systems? These are systems that can chat like a human and perform other tasks such as writing computer code. Examples include ChatGPT by OpenAI, Bard from Google, and Bing Chat from Microsoft.

• Every day

• Several times a week

• Once a week

• 1-3 times a month

• I have tried them but did not continue using them.

• I'm aware of them, but I have not used them.

• I've never heard of them.

Why benefits and tradeoffs are included in policy questions?

Questions 5, 6, and 7 of the survey dive into a number of policy questions. These are presented as “Ideas” since their proposals. I had several conversations with NORC about the best way to structure these questions. Initially, the questions presented the policy, but cautioned that there were tradeoffs. Here’s what question 5d initially looked like:

AI-generated content should be clearly labeled as such to inform the audience. However, this might reduce the perceived value or authenticity of the content, potentially leading to less engagement or interest from users.

The wording of this question is far too complex. This gets a grade of ‘D’ from Readable. As part of the language simplification I also decided to move the tradeoff to a separate line. Making the tradeoff more explicit and separate from the idea was inspired by a reviewer from the Centre for the Governance of AI. The point of the clarification is that the question is asking about agreement with the Idea itself, not the tradeoff.

The next version of the question looked like this:

Idea: Content created by AI should have a clear label saying it's AI-generated. AI-generated content includes images, videos, stories, and music.

Tradeoff: People might engage with content less if they know it's made by AI even if the content is useful or entertaining. That's because it could seem less genuine or appealing.

I also added an explicit direction to the question setup: “Keep in mind, your opinion is about the idea itself, not its tradeoff.” Further, I was able to eliminate two response options, one from Question X and one from Question X to …

Nonetheless, there was then concern from NORC that the tradeoff was explicit, but the benefit was not. A critique I agree with. There was also a suggestion to change the font type of the tradeoff to further emphasis it was for context only. I made these updates, which resulted in the following final format.

Idea: Content created by AI should have a clear label saying it's AI-generated. AI-generated content includes images, videos, stories, and music.

Benefit: This could enable users to understand content origins and make informed decisions.

Tradeoff: People might engage with content less if they know it's made by AI even if the content is useful or entertaining. That's because it could seem less genuine or appealing.

But concern then shifted to question complexity. Providing both a benefit and a tradeoff for a series of grid questions does add cognitive burden and increase the length of the survey. It could also be that some people read the tradeoff and benefit while others ignore it, creating a disparity in content.

Nonetheless I decided to keep both the benefit and the tradeoff. I felt it was important given prior research I had seen on how priming respondents to think about tradeoffs can substantially shift support for a policy. As one example support for federal student loan forgiveness was shown to change dramatically when respondents were presented with a set of tradeoffs compared to asking about abstract support without a tradeoff present. (See results below).

Tradeoffs are part of life and I wanted to include both a benefit and a tradeoff to represent that fact. I worried that without presenting tradeoffs there would be an abundance of agreement, especially given anxiety and media attention over Generative AI. I feared that for a substantial portion of respondents, any policy option that presented some way of restricting, regulating, or governing Generative AI might be met with an immediate “yes.” I wanted to interrupt the tendency by encouraging respondents to consider the benefits and tradeoffs of the Idea.

Why was forced choice used rather than “Select all that apply”?

Several questions presented a set of options and asked respondents to chose a single option. For example, here’s Question 9:

Which job area do you think will benefit the most from Generative AI in the next 3 years? For each job area we've given some example jobs, but there might be other jobs in the job area.

• Creative jobs (Artists, Writers, Designers)

• Media jobs (Game designers, Game programmers, 3D artists)

• Performing arts (Movie and TV actors)

• Finance jobs (Accountants, Stock analysts, Traders)

• Tech jobs (Programmers, Data analysts, Engineers, IT)

• Teaching jobs (Teachers, Course creators)

• Legal jobs (Lawyers, Legal assistants)

• Health jobs (Doctors, Nurses)

• Sales and Marketing jobs

• Other (please specify)

• I don’t know

There with NORC about whether this question should instead be of the form, “Select all that apply.” The recommendation from NORC was that while choose-all-that-apply style questions have their place, forced choice is usually better. This aligns with Stantcheva’s advice as well:

For questions where there can be multiple answer options selected, you have to decide between using a “forced-choice” or a “check-all-that-apply” format. Forced-choice questions ask item by item and require respondents to judge all items presented independently. Check-all-that-apply formats list all options simultaneously and ask respondents to select some of the items presented (Smyth et al., 2006). Forced-choice questions generally lead to more items being selected and respondents thinking more carefully about the answer options. As discussed below, forced-choice questions will also circumvent the problem of order effects in the answer options, whereby respondents may be tempted just to select one of the first answers and move on without considering every choice. If you can, try to convert your “check-all-that-apply” questions into individual forced-choice questions.

Select-all-that-apply questions tend to lead to a lot of acquiescence. Consider a recent survey of misinformation experts administered by Harvard. The question asked respondents about how strongly they agree that a set of different factors cause people to believe misinformation. Overall, the survey is insightful for sure, and we do learn something about the experts’ opinions. Still, I wonder if this question would’ve been more useful presented as a forced choice, about what the main reason people believe misinformation. As it stands it’s

We know from the Piraña Principle that not all of these factors can be primary drivers of misinformation belief. Response patterns of the top five factors are almost identical and so we learn very little about how experts view the importance of these factors relative to one another. Should we tackle partisanship first or lack of trust in institutions?